AI Paper Atlas

Research map / Large language models / Agents, code and math

Retro-Expert: Collaborative Reasoning for Interpretable Retrosynthesis

Xinyi Li, Sai Wang, Yutian Lin, Yu Wu

ICML 2026 · Large language models › Agents, code and math · approach: LLMs for scientific discovery · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
RL with KGPO instead of SFT with CoT (ablation)Qwen2.5-7B-Instruct with decision space · USPTO-50KTop-1 accuracy (%): 48.6 -> 75.7Improved
replace standard GRPO with KGPO advantage rectification (ablation)Retro-Expert RL · USPTO-50KTop-1 accuracy (%): 66.8 -> 75.7Improved
add reaction-related reward rreact to outcome reward (ablation)Chem-RM · USPTO-50KTop-1 accuracy (%): 66.8 -> 72.3Improved
add structure-related reward rstruct (ablation)Chem-RM · USPTO-50KTop-1 accuracy (%): 72.3 -> 75.7Improved
add decision space from specialized models to general LLMs (ablation)GPT-4o · USPTO-50KTop-1 accuracy (%): 0.30 -> 32.85Improved
add decision space to specialized chemical LLMs (ablation)T5Chem and nach0 · USPTO-50KTop-1 accuracy (%): 43.64 -> 27.88 (T5Chem); 30.52 -> 14.60 (nach0)Hurt
add decision space to SFT-trained Qwen2.5-7B-Instruct (ablation)Qwen2.5-7B-Instruct (SFT) · USPTO-50KTop-1 accuracy (%): 43.62 -> 58.92Improved
use Retro-Expert on top of specialized retrosynthesis modelsGraph2Edits, UAlign, LocalRetro, GraphRetro, GLN, Retroformer, MHNReact · USPTO-50KTop-1 accuracy (%): Graph2Edits 67.2 -> 75.7; MHNReact 51.8 -> 60.9Mixed
use Retro-Expert instead of the best baselineQwen2.5-7B-Instruct with KGPO · USPTO-FULLTop-1 accuracy (%): 49.7 -> 58.4Improved
use Retro-Expert instead of DeepSeek-R1 and base Qwen2.5-7Bout-of-distribution reactions · ChemBenchTop-1 accuracy (%): 29.33 -> 65.67Improved
replace GRPO with KGPO (ablation)candidate rejection behavior · USPTO-50K testrejection rate of wrong candidates (%): 0.22 -> 39.12Improved
use KGPO instead of GRPO (ablation)reasoning quality · USPTO-50KGPT-4o Mechanism Accuracy (1-5): 3.37 -> 4.41Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy